refactor: 统一硬件相关头文件引用

将分散在各文件中的CUDA/HIP/MUSA硬件相关头文件引用统一到vendors目录下的对应头文件中，提高代码可维护性。移除重复的头文件引用，优化构建配置。

csrc/quantization/cutlass_w4a8/get_group_starts.cuh

@@ -1,9 +1,7 @@
 // see csrc/quantization/w8a8/cutlass/moe/get_group_starts.cuh
 #pragma once
 
-#include <cuda.h>
-#include <torch/all.h>
-#include <c10/cuda/CUDAStream.h>
+#include "../../vendors/functions.h"
 
 #include "core/scalar_type.hpp"
 #include "cutlass/bfloat16.h"

csrc/quantization/cutlass_w4a8/w4a8_grouped_mm_entry.cu

@@ -14,9 +14,8 @@
 #include "cutlass/util/mixed_dtype_utils.hpp"
 
 // vllm includes
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-#include <torch/all.h>
+#include "../../vendors/functions.h"
+
 #include "cutlass_extensions/torch_utils.hpp"
 #include "cutlass_extensions/common.hpp"
 

csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu

@@ -3,9 +3,10 @@
 //   https://github.com/NVIDIA/cutlass/blob/main/examples/55_hopper_mixed_dtype_gemm/55_hopper_int4_fp8_gemm.cu
 //
 
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-#include <torch/all.h>
+
+#include "../../vendors/functions.h"
+
+
 #include "cutlass_extensions/torch_utils.hpp"
 #include "w4a8_utils.cuh"
 
@@ -26,7 +27,6 @@
 #include "cutlass_extensions/common.hpp"
 #include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
 
-#include <cuda_runtime.h>
 
 namespace vllm::cutlass_w4a8 {
 

csrc/quantization/cutlass_w4a8/w4a8_utils.cu

@@ -1,7 +1,10 @@
 #include "w4a8_utils.cuh"
 
+
+#include "../../vendors/functions.h"
+
+
 #include <array>
-#include <cuda_runtime.h>
 #include <cstdio>
 
 namespace vllm::cutlass_w4a8_utils {